The present application relates to semiconductors, and more specifically, to techniques for forming resistive memory element semiconductor structures. Arrays of trainable resistive crosspoint devices, referred to herein as resistive processing units (RPUs), may be used to perform vector-matrix multiplication in an analog domain in a parallel fashion.